Model is mathematical representation of real world phenomenon, reality
with the degree of accuracy required depending on the application
Normative Approach to Modelling
Modelling in the physical and social sciences o Normative means a theory of how something should be done. Social sciences address the questions of human behavior, controlled experiments are essential to physical sciences which are difficult to do in social science, and it has implicit conviction that models are capable of describing reality very accurately. o Statistical inference means diligent experiment and study to refute or validate any assumptions. Exploratory Data Analysis o Starting point to collect and do analysis of data. Exploratory data analysis looks for pattern, often done graphically. This is to find a proper formula or model. o 3 reasons for choosing a model in a form: statistical evidence, a matter of common sense, have successfully used the data before. Model Calibration o It means estimating the uncertain parameters in a model o Usually done by Maximum Likelihood Estimation (frequentist or classical) or by Bayesian Approach. o Monte Carlo Markov Chain has been a recent breakthrough o Our ability to estimate the data is limited by the noise in the data. Fit to evidence o Whether this model is correct or useful o Do back-testing, that is to check whether the model mean and variances are comparable to the data used o Do out of sample or blend test valuation, that is to set aside recent data during calibration process and use it for testing o Generate future projections and compare with data. o Longitudinal analysis follows the progress of one simulated path over time; cross sectional analysis looks at a fixed future time horizon across many simulated outcomes. Hypothesis Testing o A statistical tool to check for fit Test Results support Test Results support H0 H1 H0 is Good Outcome = 1- Type II Error = a true a H1 is Type I Error = b Good Outcome = 1 true b o Most powerful test is given a value of a, one has the smallest possible value for b Parsimony o Adding more parameters might improve fit, but this violate the principle of parsimony -> using only as many parameters as necessary to fit a model o When you use too many parameters you are simply modelling the samples to perfection, this idea is incorrect because the parameters should represent the population. o To account for this, penalty based on AIC and BIC. Fit to theory o Chosen model structure could be justified by reference to theory, instead of rigorously testing all of them formally. o Some models can be self-referential, that is the use of models could actually change the behavior of the phenomenon being modeled. Computer technology development o Translating formula into programming codes is different thing, and commonly exposed to mistakes. o Here are some proactive way to look for errors: Numerical checks to validate algebra, sense check model outputs (Explainable movements), re test software functionality, documents the tests performed to avoid repeated testing effort, mistakes may not be down to your codes o Many started coding to early on unstructured basis, condemned to months of tracking down and fixing bugs. To deal with this situation we have defensive programming. Using models for projections o Statistical textbook often emphasizes the value of accompanying every estimate with a measure of its uncertainty o There may be error when back testing and it falls beyond the CI. These errors may be caused by: process error, parameter error, model error, calibration error, survivorship bias, and operational errors. Bootstrapping o Labor intensive but powerful test. Computational model classification o Deterministic Applies a set of formula without introducing any randomness A probabilistic setting has been transformed into a deterministic model by replacing all random items by their expected value A model office is a model of the financial statements and cash flows of an insurance operation o Stochastic A model with explicit reference to different probabilities Limitations on Normative Approach Practical difficulties o There are many interpretations of normative, and what was chosen may just be one of many. o Statistical inference might show support to a small or poorly understood data sets, which might have just been the only available source. o A better model might be omitted due to its little impact on results Theoretical ambiguities o When comparing models in hypothesis test, you could simply compare with an implausible model to simply validate your choice o Within a 95% CI, testing for 20 models will leave 1 model being the better of existing, quickly overlooking the 19 already compared. o It requires skills and judgment to construct a proper hypothesis test o Type III error of having to pick one of the neither correct HO nor H1
Expecting the unexpected
o To stress test against past scenarios, this is difficult when you dont have the experience, and may be overstated if the event is in the data. Techniques for estimating such tail event: Exposure based analysis looks at the drivers of potential losses even when no claims have been made An analysis of extreme but painful outcomes may make use of data on near misses To extrapolate data with rare events
Commercial modelling The role of modelling within the actuarial control cycle Costs of models and data Robustness Governance and control Models for advocacy Models and markets Disclosure